| Popis: |
With the increasing prevalence of digital multimedia devices and the growing reliance on compression and wireless data transmission, evaluating image quality remains a persistent challenge. This study addresses the limitations of image quality assessment stemming from the expense of data annotation and the scarcity of labeled training datasets. Leveraging visual representation learning, our approach involves training a deep Convolutional Neural Network on a large image dataset generated by simulating 165 distortion scenarios across 150,000 images, resulting in 24.75 million distorted images. These distortions are labeled using an ensemble of full-reference quality assessment models. The trained model undergoes fine-tuning on diverse datasets, including TID2013, Kadid-10K, KonIQ-10K, and BIQ2021, encompassing both simulated and authentic distortions. The fine-tuning process achieves state-of-the-art image quality assessment performance, yielding Spearman’s correlation coefficients of 0.921, 0.893, 0.884, and 0.793, respectively, for the four datasets. Comparative analysis with an ImageNet pre-trained model demonstrates superior performance in terms of Pearson and Spearman’s correlations, achieving validation criteria with fewer epochs. These findings contribute to the advancement of IQA, offering a promising approach for robust and accurate quality prediction in various applications. |
